Local voice transcription

Talk to the AI Assistant instead of typing. With local voice transcription, Pinegrow records your request and a Whisper speech-to-text model turns it into text on your computer. The recording is not uploaded for transcription. Only the transcribed text is sent to your AI provider, as a normal prompt.

Before you start

• Pinegrow Desktop. Local transcription is part of Pinegrow 10 Desktop on macOS (Apple silicon and Intel), Windows (x64) and Linux (x64). Pinegrow Online can only use a provider for transcription.
• A configured AI Assistant. The transcribed prompt runs on your usual AI provider and model.
• A microphone and disk space for a model, from 81 MiB to 1.2 GiB.

Find voice input

The microphone icon is at the right edge of the AI Assistant prompt box.

Voice settings are in AI Settings. Open it with the AI Settings button (the tools icon at the right of the Send button, below the model badge), or choose Support & Settings › AI Settings….

Set up local transcription

1. Open AI Settings and scroll to Voice transcription.
2. Set Source to Local on this computer, the default in Pinegrow Desktop. The other options, such as OpenAI or OpenRouter, upload the recording to that provider.
3. Choose the Local language. Detect automatically is the default. Pick your language to skip detection.
4. Choose the Local model.
5. If the button below the model says Not installed · Download…, click it and confirm Download. You can also skip this step: Pinegrow asks Download voice transcription model? the first time you record.
6. Click Save.

Choose a model

Pinegrow offers four Whisper models. Larger models are usually more accurate, but they use more disk space and memory and take longer to transcribe.

• Whisper Base Q8, Compact (81 MiB): the smallest download.
• Whisper Small Q8, Balanced (257 MiB): the safer choice on computers with less memory or without GPU acceleration.
• Whisper Large v3 Turbo Q8, Accurate (845 MiB): the default model.
• Whisper Large v3 Q6, Maximum accuracy (1.2 GiB): the largest model, and slower than Large v3 Turbo.

Pinegrow downloads these models from Hugging Face. It checks each download by its exact size and SHA-256 checksum, then stores it in Pinegrow's application data folder, under pine-cone/transcription/models. See Pinegrow Data folder and protected credentials.

When a model is installed, its button reads Installed (size) · Remove…. Click it to delete the model and free the disk space.

Use your own GGUF model

Choose Custom GGUF model… and select a .gguf speech-to-text model file in Custom model file. Pinegrow uses the file where it is and never deletes it.

Pinegrow only checks that the file has the .gguf extension and can be read. It does not verify a checksum or guarantee that the model works, and automatic language detection may not be available. Use models from sources you trust. See the models documented for transcribe.cpp.

Dictate a request

1. Click the microphone icon in the prompt box. The prompt box is replaced by the recorder.
2. Click the recorder to start recording, or press and hold Space. Pinegrow first shows Preparing local transcription… while it checks and loads the model, then Requesting microphone access…. The first time, your system may ask you to allow Pinegrow to use the microphone.
3. Speak your request. Bars show that Pinegrow is recording.
4. Click the recorder again to stop. If you are holding Space, release it.

With a provider as the source, the recorder shows a Provider transcription notice about the upload. With local transcription, there is no notice.

After you stop, the recorder shows Transcribing… (click to cancel). Then AI Assistant runs the text right away, as a follow-up to the current task when possible. Say "new task" at the start to start a new task. Pinegrow tells the assistant the prompt came from speech, so it corrects likely transcription mistakes, or asks what you meant and offers choices.

To go back to typing, click the crossed-out microphone icon in the recorder.

What happens to the recording

• The recording is kept in memory while you speak.
• For transcription, Pinegrow writes a temporary copy to your system's temporary folder. A separate helper process transcribes it with the selected model. The temporary copy is deleted as soon as transcription ends. If Pinegrow quits before that, leftover copies older than a day are removed the next time it starts.
• Only the transcribed text is sent to your AI provider, together with the usual context of an AI Assistant task.
• The model stays loaded for 5 minutes after you use it, then Pinegrow unloads it to free memory.

Limits

• One recording can be up to 10 minutes long. Pinegrow stops recording automatically at 10 minutes.
• For a very long recording, Pinegrow asks Are you sure about this long recording? Choose Transcribe and run the command to continue, or Delete it to discard the recording.
• Cantonese needs a Large v3 model.
• Wait until the assistant finishes the current task before you dictate. If a task is still running when transcription finishes, Pinegrow shows Please wait for the current task to complete… and the dictated request is not run.

Troubleshooting

• No microphone icon. The selected transcription source isn't available. Check Voice transcription in AI Settings. A provider source needs an active API key for that provider.
• Unable to record. Pinegrow could not access the microphone. Allow microphone access for Pinegrow in your system's privacy settings. On macOS, this is in System Settings › Privacy & Security › Microphone. Then try again.
• No speech was recorded or No speech was recognized. Check that your microphone is connected and selected as the input device in your system settings, then record again.
• Unable to transcribe audio. The message includes the reason, for example that the model isn't available. Open AI Settings, check the model button, and download the model again if needed.
• Incomplete download · Remove… A download was interrupted. Click the button to remove the partial files, then download the model again.
• Corrupt model removed · Download… The model failed verification and was deleted. Click the button to download it again.
• Status unavailable · Retry… Click the button to check the model again.
• The model button says Unavailable. Local transcription isn't supported on this operating system or processor. Choose a provider as the Source instead.
• Custom model errors. Pinegrow reports when the file doesn't exist, is empty, can't be read or isn't a .gguf file. If the model still fails, use one of the Pinegrow models.

Related

• Pinegrow AI Assistant
• Review AI changes
• Pinegrow Data folder and protected credentials